Deep-Learning based Scatter Correction in Digital Radiography

Abstract

X-ray scattering significantly reduces the contrast resolution of the image in digital chest radiography. The conventional strategy for reducing scattered radiation is the use of anti-scatter grids which, while improving image quality, increase the radiation dose absorbed by the patient and pose geometrical restrictions in non-standard techniques such as portable radiography. In this work, we propose and evaluate two approaches for scatter correction based on deep learning, which adopt a U-net convolutional neural network architecture. An indirect method, based on the estimation of the scatter map which is then subtracted from the original image, and a direct method, based on the estimation of final corrected image directly. Due to the lack of real acquisitions with and without anti-scatter grid, Monte Carlo simulations were performed to generate the dataset. This study demonstrates the potential of the DL methods to remove the scattered radiation with an error of around 5%.